Deep Syntactic Processing by Combining Shallow Methods

Deep Syntactic Processing by Combining Shallow Methods P´eter Dienes and Amit Dubey Department of Computational Linguistics Saarland University PO Box 15 11 50 66041 Saarbr¨ucken, Germany {dienes,adubey}@coli.uni-sb.de Abstract We present a novel approach for find- ing discontinuities that outperforms pre- viously published results on this task. Rather than using a deeper grammar for- malism, our system combines a simple un- lexicalized PCFG parser with a shallow pre-processor. This pre-processor, which we call a trace tagger, does surprisingly well on detecting where discontinuities can occur without using phase structure information. 1 Introduction In this paper, we explore a novel approach for find- ing long-distance dependencies. In particular, we detect such dependencies, or discontinuities, in a two-step process: (i) a conceptually simple shal- low tagger looks for sites of discontinuties as a pre- processing step, before parsing; (ii) the parser then finds the dependent constituent (antecedent). Clearly, information about long-distance relation- ships is vital for semantic interpretation. However, such constructions prove to be difficult for stochas- tic parsers (Collins et al., 1999) and they either avoid tackling the problem (Charniak, 2000; Bod, 2003) or only deal with a subset of the problematic cases (Collins, 1997). Johnson (2002) proposes an algorithm that is able to find long-distance dependencies, as a post- processing step, after parsing. Although this algo- rithm fares well, it faces the problem that stochastic parsers not designed to capture non-local dependen- cies may get confused when parsing a sentence with discontinuities. However, the approach presented here is not susceptible to this shortcoming as it finds discontinuties before parsing. Overall, we present three primary contributions. First, we extend the mechanism of adding gap vari- ables for nodes dominating a site of discontinu- ity (Collins, 1997). This approach allows even a context-free parser to reliably recover antecedents, given prior information about where discontinuities occur. Second, we introduce a simple yet novel finite-state tagger that gives exactly this information to the parser. Finally, we show that the combina- tion of the finite-state mechanism, the parser, and our new method for antecedent recovery can com- petently analyze discontinuities. The overall organization of the paper is as fol- lows. First, Section 2 sketches the material we use for the experiments in the paper. In Section 3, we propose a modification to a simple PCFG parser that allows it to reliably find antecedents if it knows the sites of long-distance dependencies. Then, in Sec- tion 4, we develop a finite-state system that gives the parser exactly that information with fairly high accu- racy. We combine the models in Section 5 to recover antecedents. Section 6 discusses related work. 2 Annotation of empty elements Different linguistic theories offer various treatments of non-local head–dependent relations (referred to by several other terms such as extraction, discon- tinuity, movement or long-distance dependencies). The underlying idea, however, is the same: extrac- tion sites are marked in the syntactic structure and this mark is connected (co-indexed) to the control- Type Freq. Example NP–NP 987 Sam was seen * WH–NP 438 the woman …

Content truncated to 500 words